Logocentrism

From the AI Research Compass wiki

Logocentrism: Ring title for extreme Language scores: language is not one modality among many but the medium in which thought itself happens.

Beliefs

Logocentrism is a ring title earned when the Language-World score falls below -0.85, the far left edge of the chart, whatever the position on the vertical axis. The name borrows a term from philosophy and wears it with pride rather than as a criticism: in the beginning was the Word. Its holders believe reasoning is fundamentally linguistic, that a chain of thought is the thinking rather than a report of it, and that perception and action matter only insofar as they can be described, labelled and reasoned about in language.

Because it is about the horizontal extreme alone, a logocentrist can be a monolith-scaler or a harness-builder; what unites them is that every other modality is translated into words before anything important happens. Its Ultra-Orthodox shade reflects that this is the reigning consensus of the language model era, defended against the rising tide of world-model and embodiment advocates.

Arguments

  • Language models trained purely on unlabelled text learned to perform many tasks zero-shot, suggesting that task competence is largely encoded in language.
  • Most progress in multimodal systems has come from attaching perception to a stronger language backbone rather than the reverse.
  • Language is compositional, abstract and shareable; it is how humans transmit knowledge across generations, which no other modality does as well.

Critiques

  • Much human and animal competence, from catching a ball to recognising a face, is not linguistic and cannot be reduced to words without loss.
  • Models that reason only in text still fail at spatial, physical and quantitative tasks that are trivial for a child with a body.
  • Treating language as the medium of all thought may simply reflect what is easiest to collect and evaluate, not what intelligence requires.

Representative works

Relationships

How to get it

Earned in the full or pro test when: Representation <−0.85.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data